A new BP network model was developed to improve the accuracy and assess the groundwater quality. For this purpose, the krill herd algorithm (KHA) was established with the optimization process of the connection weights and thresholds of the BP neural network. Totally 15 farms were selected to evaluate the groundwater quality and identify the main causes of groundwater pollution in Jiansanjiang Administration. In addition, to verify the applicability of the model, the distinction degree method and the theory of serial number summation were used to analyze the reliability and stability of KHA-BP model, PSO-BP model and BP model, respectively. The results exhibited a good agreement of groundwater quality in each farm and there was a certain spatial distribution pattern such as the water quality of grade I was mainly concentrated in the southwest position, grade Ⅱ was distributed in the north and south, while the grade Ⅲ was located in the midwest and mideast of the administration. Fe, Mn, CODMn, NH3N and NO-3N were the main factors caused groundwater pollution. Fe and Mn were local primary hazard but excessive amounts of CODMn, NH3N and NO-3N were mainly related to use of a large number of fertilizers and pesticides. The distinction degree of KHA-BP was 1.1070 and Spearman’s rank coefficient was 0.9286, which was better than those of PSO-BP and BP. In conclusion, this research could provide a scientific basis for the comprehensive management of groundwater resources and construction of water ecological civilization in the core areas of food production.